Mechanistic Accountability in Hybrid Physiologically Based Pharmacokinetic and Machine-Learning Models for Drug Development Download PDF

Hybrid physiologically based pharmacokinetic and machine-learning models offer a potentially valuable means of combining physiological organization with data-adaptive estimation. Their scientific interpretation is difficult, however, because predictive performance does not reveal whether a learned component represents a biological quantity, compensates for structural error, exploits dataset-specific associations, or remains reliable under extrapolation. This article develops mechanistic accountability as an original modeling principle for hybrid PBPK–machine-learning systems in drug development. Mechanistic accountability is defined as a traceable, context-specific justification linking each learned quantity to its data provenance, semantic meaning, interface with PBPK structure, effect on mechanistic states, identifiability, propagated uncertainty, biological plausibility, validation evidence, comparator performance, and permissible decision influence. The proposed construct distinguishes PBPK structure from physiological truth, learned inputs from measured parameters, prediction from mechanistic explanation, explainability from accountability, and model confidence from calibrated uncertainty. It further separates four possible machine-learning roles: estimating PBPK inputs, learning parameter or covariate functions, correcting mechanistic discrepancy, and directly predicting pharmacokinetic outputs. Each role creates different evidentiary obligations and failure modes. The article argues that hybrid-model evaluation should integrate software and interface verification, identifiability analysis, uncertainty propagation, biological challenge testing, external predictive assessment, and comparison with simpler alternatives. Qualification should remain bounded to a defined drug-development question, population, scenario, model influence, and consequence of error. The proposed framework is conceptual rather than empirically validated and does not establish regulatory acceptance, clinical utility, or deployment readiness. Its principal implication is that hybrid models should be judged not by a single performance measure but by whether their learned and mechanistic elements can sustain a transparent, proportionate, and scientifically defensible claim.